Improving TCP PerformanceDuring the Intra LTE Handover

D. Pacifico, M. Pacifico, C. Fischione, H. Hjalrmasson, K. H. JohanssonSchool of Electrical Engineering

KTH Royal Institute of Technology, Stockholm, SwedenEmail: {davidep|pacifico|carlofi|hjalmars|kallej}@ee.kth.se

Abstract—Ensuring a seamless connection when users aremoving across radio cells is essential to guarantee a highcommunication quality. In this paper, performance of TCPduring the handover in a Long Term Evolution (LTE) networkis investigated. Specifically, mobile users with high bit ratesTCP services are considered, and the impacts of the intra LTEhandover over their perceived throughput are studied. Due tothe mobility of the users across radio cells, the high bandwidthrequired, and possible network congestions, it is shown thatthe handover may cause sudden degradation of the quality ofthe communication if the process is not correctly controlled. Toalleviate these problems, three solutions are proposed: fast pathswitch, handover prediction, and active queue management. Thefirst two solutions avoids excessive delay in the packet deliveryduring the handover, whereas the second solution acts at thetransport network with an active queue management. Simulationresults, obtained by an extension of the ns-2 simulator, show thatthe proposed solutions present advantages, and that the handoverprediction used with the active queue management increases TCPperformance significantly.

I. INTRODUCTION

The recent increase of mobile data usage and the emergenceof new applications, such as multimedia online gaming, mobileTV, Web 2.0, and streaming contents, have motivated thedevelopment of LTE, the Long Term Evolution of the UMTSterrestrial radio access network [1]. LTE supports a highthroughput with low latency by an IP based transport network.TCP, the Transmission Control Protocol, is expected to playa significant role in LTE, given its high popularity on IPnetworks.

To achieve a seamless mobility across radio cells, LTEimplements a hard handover algorithm, which moves a mobileuser from one base station serving a radio cell to anotherone. However, considering the high bandwidth required bythe users, the handover may cause a sudden degradation ofthe throughput of the TCP connections. Such a problem can befurther exacerbated if the wired transport network is congested.Since the handovers may occur frequently for LTE networks,we conclude that the quality of service perceived by the usersthat are using TCP connections may be low.

There is a large literature on research for improving TCPover wireless links [2]. The main approaches proposed can

The work of the authors was supported by the EU projects FeedNetBack,the Swedish Research Council, the Swedish Strategic Research Foundation,and the Swedish Governmental Agency for Innovation System.

be grouped into three categories: end-to-end solutions (e.g.,Eifel [3]), physical layer solutions (e.g., [4], [5]), and cross-layer solutions (e.g., [6]–[9]). However, the high bit rateoffered by LTE, which is comparable to or even larger thanwired links, alleviates substantially the problems of TCP overmore traditional wireless networks investigated in the liter-ature. Performance of TCP over LTE is affected mostly bythe links of the wired network and total bandwidth availableat the serving base station. In [10], an end-to-end solutionto improve TCP during the handover has been proposed, butthe drawback is that it requires modification of existing TCPprotocols. The impact of LTE handover on user connection hasbeen considered in [11], where some scheduler modificationsto improve the user throughput are suggested. However, thecore network status is not considered, where the network loadplays a significant role on performance. As a consequence,scheduler modifications may have limited effect on the TCPthroughput. In [12] a user data forwarding during the handoveris investigated to increase the throughput.

In this paper we investigate TCP performance during theintra LTE handover procedure in a complete network scenario.We consider the data forwarding technique studied in [12]and propose some mechanisms to increase performance sig-nificantly. We show that TCP suffers of throughput degra-dation during the handover. Consequently, we propose someimprovements of different complexity, namely: two forwardingavoidance algorithms (a simple one, and a second one that usesa predictor) and active queue management. These solutionsrequire only a slight modification of the existing LTE handoverprocedure, and act mostly at the network layer. Differentlyfrom previous studies, we improve performance of TCP duringLTE handover by acting inside the transport and not at theaccess network. We evaluate performance of the proposedsolutions by an extension of the ns-2 simulator. The forwardingavoidance algorithms with the predictor shows the largestimprovement, but at the cost of a slightly increased complexityof the handover algorithm.

The remainder of the paper is organized as follows: InSection II, we describe in detail the problem we are addressing.In Section III background information on TCP and handover isgiven. A performance analysis of TCP over LTE is developedin Section IV. Improvements are proposed in Section V. InSection VI, performance of these solutions are investigated.

SourceCell

eNodeB

Router

Server

ME

Target Cell

Gatewayref ME

Fig. 1. System Model.

Finally, in Section VII, conclusions and future perspectivesare given.

II. PROBLEM FORMULATION

We consider a system scenario with two cells served bytwo different evolved Node B (eNodeB), a variable number ofMobile Equipments (MEs), divided among the source and thetarget cell (see Fig. 1).

An ME is doing an intra LTE handover from the firsteNodeB to the second one. Let us denote such an ME by refME. ref ME is using a TCP connection to download data froma server. The remaining MEs are competitors for ref ME forthe usage of the available bandwidth. A core-network routerconnects the two eNodeB to a server pool in which each serverworks as traffic source for a number of MEs. Let SeNodeBand TeNodeB be the Source and Target eNodeB for ref ME,respectively.

During the handover procedure, packets waiting in theSeNodeB to be sent to the ref ME are forwarded from theSeNodeB to the TeNodeB via a core-network router. Thismechanism has negative effects for the TCP round trip timeof ref ME. The forwarding may cause an increase of delay inthe packet delivery, especially when the network is congested.This could cause a retransmission timeout expiration of TCPand the congestion window could drop to 1 segment. As aresult, the ref ME could experience a dramatic drop of thethroughput during the handover. Therefore, in this paper wetackle this problem and propose some control mechanisms tocounteract such an adverse situation.

In the next section, we give some background informationneeded for the evaluation of performance loss and the designof these control mechanisms.

III. BACKGROUND

In this Section we give an overview of LTE, TCP, andhandover.

A. TCP over LTE

LTE is expected to improve substantially end-user through-put, cell capacity, and transmission latency. Given the popu-larity of the Transmission Control Protocol TCP, and Internet

Protocol (IP) for carrying all types of traffic, LTE supports TCPand IP-based traffic with end-to-end quality of service.

TCP [13] is a reliable transport protocol that uses a con-gestion window cwnd to send packets. The TCP reliability isachieved using ARQ mechanism based on positive acknowl-edgments. TCP packets are cumulatively acknowledged whenthey arrive in sequence; out of sequence packets cause thegeneration of duplicate acknowledgments. TCP sender detectsa packet loss either when multiple duplicate acknowledgments(3 is the default value) arrive or when a retransmission timeout(RTO) expires. The RTO value is calculated dynamically basedon RTT measurements.

LTE assumes that the end-to-end TCP protocol governsthe congestion control and adapts to the varying networkconditions and handle packet loss. The LTE system supportsquality of service mechanisms over radio and in the transportnetwork, but no flow control mechanisms are supported. Thisimplies that packet buffering/dropping for TCP could occurduring congestion in any node of the core network (e.g., inthe base station or a core-network router).

In LTE, the connection from a server to the MEs crosses thecore network exploiting an IP tunnel [14], which goes betweenthe serving gateway and eNodeB, and is based on the GPRSTunneling Protocol (GTP). Due to this tunneling protocol, anexplicit congestion signaling is not allowed even if the userconnection were able to handle it. The IP header handled bythe core network router belongs to the tunnel application ratherthan to the user, and the signaling will be trashed once thepacket has reached the end of the tunnel. We see in the nextsubsection that such a tunneling protocol puts some constraintsto the improvement of TCP performance during LTE handover.

B. Handover

During a handover, the TCP protocol endpoints must bemoved from the SeNodeB to the TeNodeB. The message chartof the procedure is reported in Fig. 2 [14]. The SeNodeBcollects the ref ME measurements about the link quality, andassists the function controlling the ref ME’s handover. Next,SeNodeB makes a decision based on measurement reports andRRM information to hand off the ref ME, and it issues ahandover request message to the TeNodeB.

After the last phase described above, an admission controlmay be performed by the TeNodeB according to the receivedquality of service information. The TeNodeB prepares thehandover and sends the handover request acknowledge to theSeNodeB. Such a message may also include information aboutRadio/Transport Network Layer (RNL/TNL) for the forwardingtunnels. When the handover request acknowledge is received,data forwarding from the SeNodeB to the TeNodeB throughthe router (recall Fig. 1) may be initiated. The SeNodeBgenerates the handover command (RRC message) towardsthe ref ME. Then, the ref ME performs a synchronizationto TeNodeB and accesses the target cell. It then sends thehandover message to indicate that the handover procedure iscompleted.

MESource

eNB

Target

eNBMME

Serving

Gateway

End Marker

0. Area Restriction Provided1. Measurement

ControlPacket Data Packet Data

UL allocation

2. Measurement

Report3. HO

decision

4. Handover

Request

6. Handover

Request Ack

5. Admission

Control

DL allocation

7. Handover

CommandDetach from old

cell and

synchronize to

new cell

Deliver buffered

and in transit

packets to target

eNB

8. SN Status

TranferData Forwarding

Buffer packets

from Source eNB9. Syncronization

10. UL Allocation + TA for UE

11. Handover

Confirm 12. Path Switch

Request 13. User Plane

Update Req

14. Switch DL

Path

15. User Plane

Update Reply16. Path Switch

Request Ack17. Relase

Resource

End Marker

Flush DL buffer,

continue

delivering in

transit packets

Data Forwarding

18. Relase

Resources

Packet Data Packet Data

Ha

nd

ove

r P

rep

ara

tio

nH

an

do

ve

r E

xe

cu

tio

nH

an

do

ve

r C

om

ple

tio

n

L3 Signaling

User Data

L1/L2 Signaling

Legend

Fig. 2. Message chart of the LTE handover procedure. The control planemessages (solid and dot-dashed arrows) and the flow of the user packets(dashed arrows) are reported [14].

The TeNodeB can now start sending data to the ref ME and,at the same time, send a path switch message to the serverMME to inform that the ref ME has changed cell. The MMEsends a user update request message to the serving gatewaywhich switches the downlink data path to the TeNodeB andsends one or more ”end marker” packets on the old path tothe SeNodeB, which can then release any resources previouslyassociated to the ref ME (steps 7 to 16 in Fig. 2).

During the handover there is a time interval when packetson the new direct path (serving gateway – TeNodeB) andpackets on the forwarding path (serving gateway – SeNodeB– core-network router – TeNodeB) may arrive in parallel atthe TeNodeB. This gives rise to a possible increase of theRTT of the forwarded packets, since they have to do twice aqueueing in the core-network router. Moreover, there is alsothe potential problem of out of order packet delivery to theref ME when it is attached to the TeNodeB. In [12], a packetreordering feature inside the TeNodeB has been proposed toavoid the out of order delivering. Nevertheless, the potentialincrease of the RTT is still present, as we investigate in thefollowing.

SStreshHO

SStresh

CWNDMAX

RTT1 RTT2RTT

CW

ND

Ideal case

Window halving

Tim

eout

UnsentWH

UnsentWH

Fig. 3. Difference between ideal case and timeout occurrence.

IV. PERFORMANCE ANALYSIS

In Section III-B we have shown that user data are for-warded from the SeNodeB to the TeNodeB via a core-networkrouter during the handover procedure. This mechanism hasdetrimental effects on the RTT, because the forwarded packetmay experiments a dramatic increase of the delay, especiallywhen the receiving network is congested. In particular, theforwarded packets increase the RTT as measured at the serverdue to two queuing phases. This first queuing is experiencedby the packet transmission from the server to the core-networkrouter, whereas the second queuing is due to the forwardingfrom the SeNodeB to the router during the handover. If thesequeuing increase the RTT significantly, the cwnd could halve,or drop its value to 1 segment at worst. In the following, weanalyze such a performance degradation in detail. In particular,we examine both cwnd halving and timeout occurrence,compared with the optimal case, where cwnd does not sufferthese problems.

Fig. 3 shows the behavior of the cwnd during the handover.We assume that the cwnd is in the steady state and we focuson the time interval between an initial instant RTT1 and a finalinstant RTT2. During such an interval we want to quantify thereduction of amount of data that is not transmitted due to thehandover. For performance analysis, we define as throughputduring the handover such an amount of data. We distinguishthree possible situations: an optimal case, a window halvingcase, and a timeout occurrence case. We analyze them in thesequel.

In the optimal case the congestion window keeps on follow-ing the steady-state course regardless of the handover event.The amount of data delivered during the period of interest is

DI =RTT2∑

i=RTT1

cwnd(i)fDI=

cwndMAX∑i=SSthresh

ifDI

=32· cwndMAX

2

(cwndMAX

2+ 1)

fDI, (1)

where cwndMAX is the maximum value assumed by thecwnd in the steady-state phase, SSthresh = cwndMAX/2according to the TCP standard, and fDI

is the probability ofthe steady-state case occurrence.

The window halving case is due to an out-of-order de-livering. It happens if the router buffer exceeds the maxi-mum length during the packet forwarding. The difference ofthroughput between the optimal case and this case is shown inFig. 3 with horizontal pattern. Specifically, we need to subtractthe horizontal patterned area UnsentWH in Fig. 3 to (1) toobtain the amount of data delivered:

DWH =(DI − UnsentWH)FDW H= DIfDW H

+

−cwndMAX∑i=cwndHO

(cwndHO − SSthreshHO) fDW H

=[DI −

cwndHO

2· (cwndMAX − cwndHO + 1)

]fDW H

,

(2)

where cwndHO is the value assumed by the cwnd at thehandover start-time, SSthreshHO = cwndHO/2, and fDW H

is the probability of the window halving case occurrence.By using such a model, the allowed values for cwndHO arethose belonging to the [SSthresh,cwndMAX] interval thatcorresponds to all real possible values.

The timeout case is the worst event that can occur duringthe handover. It is a consequence of spurious timeout dueto the high jitter increase, which is exacerbated by the dataforwarding. Fig. 3 shows that in such a case we have more lossthan the previous two cases. The unsent amount of data canbe divided in two different contributions: UnsentWH , whichis due to the cwnd halving, and UnsentSS (diagonal pattern)which is due to the slow-start phase. The throughput can bewritten as

DSS =(DI − UnsentWH − UnsentSS)fDSS

= [DWH − k · (cwndMAX − cwndHO + 1) +

−k−1∑i=0

(cwndHO

2− k + i− 2i

)]fDSS

. (3)

where fDSSis the probability of timeout occurrence. The

UnsentSS quantity depends on k, which represents the totalnumber of RTT in which the slow-start phase is on duty. TheSSthreshHO belongs to the following interval

SSthreshHO ∈(

SSthresh

2, SSthresh

), (4)

where SSthresh =cwndMAX

2and k are

k = dlog2 SSthreshHOe+ 1. (5)

The theoretical evaluation of the probabilities of the optimalcase, the window halving case, and the timeout case isdifficult. Hence, we resorted to a simulation approach. InSubsection VI-B, we show that fDI

≤ fDW H≤ fDSS

. Inother words, the timeout case has the highest probability. Forperformance comparison, in Fig. 4 we evaluate the normalizedthroughput by setting fDI

= fDW H= fDSS

. Fig. 4 showsthe throughput for two handover situations as achieved by (2)(3) as function of cwndMAX parameter and normalized with

40 50 60 70 80 90 100 110 12050

55

60

65

70

75

80

85

90

95

100

CWNDMAX[packets]

Nor

mal

ized

Thr

ough

put[

%]

Timeout Occurrence

Window Halving

Fig. 4. Two possible averages of the throughput given by (2) (3) as functionof cwndMAX parameter and normalized with respect to (1). The continuouslines represent the average of the normalized throughput when the handoveroccurs when the the congestion window is at the maximum cwndMAX (seeFig. 3). The lower circled and upper crossed lines assume that the handoveroccurs in the middle of the interval (SSthresh,cwndMAX) (see Fig. 3).

HO decisionHandover Request

Handover RequestAck

Admission Control

DL allocationHandover Command Path Switch Request

User Plane UpdateReq

Switch DL PathEnd Marker

ME SourceeNB

TargeteNB MME Serving

Gateway

L3 Signaling

User Data

L1/L2 Signaling

Legend

Fig. 5. Fast path switch modification to the handover message chart.

respect to (1). We observe a substantial performance degrada-tion, which suggests the need of improving solutions for TCPduring LTE handover.

V. IMPROVING HANDOVER

In this section, we propose some methods to enhanceTCP performance during handover. The solutions are groupedinto two categories: Forwarding avoidance, and active queuemanagement. Details follow in the sequel.

A. Forwarding Avoidance

The basic idea is to avoid the packet forwarding from theSeNodeB to the core-network router and prevent that the RTTincreases so much that a spurious timeout occurs. This canbe accomplished in two ways: by a modification of the pathswitch time, or a prediction of the handover. In the current LTEhandover procedure (see Fig. 2) the path switch command issent by the TeNodeB only after it has received the handoverconfirm message from the ref ME when it is attached to theTeNodeB. The solutions proposed in the following subsectionsrequire modifications of such a handover procedure in that theresource allocation on the TeNodeB is done after a requestmade by the network and not by the ref ME.

1) Fast Path switch: This solutions consists in sending thepath switch command to the serving gateway immediately afterthe handover command is sent by the SeNodeB, instead of

HO predictionHandover Request

Handover RequestAck

Admission Control

DL allocationHandover Command

Path Switch RequestUser Plane Update

ReqEnd Marker

1. MeasurementControl

Packet DataUL allocation

2. MeasurementReport

Nec

essa

ryP

redi

ctio

ntim

ein

terv

alME Source

eNBTargeteNB MME Serving

Gateway

Switch DL Path

L3 Signaling

User Data

L1/L2 Signaling

Legend

Fig. 6. Modification to the handover message chart for the handoverprediction.

making the path switch upon the reception of the handoverconfirm message by the TeNodeB. Indeed, in the current LTEhandover algorithm, even though the handover command issent, the serving gateway keeps sending packets to the SeN-odeB during the whole duration of the handover execution, asdiscussed in the end of Subsection III-B. Hence, the objectiveis to reduce the number of packets that do a double queuingin the core-network router, which translates in a reduction ofthe potential RTT increase. The modification of the first partof the message chart required by the fast path switch solutionis reported in Fig. 5.

2) Handover prediction: The previous proposal has theadvantage of simplicity. However, there can be packets doinga double queuing. To overcome this limitation, it is possi-ble to further anticipate the path switch command of theprevious solution by a prediction of the handover start-time.The prediction can be done easily up to the coherence time,which spans over hundred of milliseconds []. By doing sucha prediction, the packets sent so far (and that are queued inthe core-network buffer) follow the path given by the server– core-network router – SeNodeB and are sent to the RefME during the time interval between the anticipated patchswitch command and the handover command. The new packetsthat are sent afterwards the anticipated path switch commandfollows the new path given by the server – core-networkrouter – TeNodeB, thus avoiding a double queuing in thecore-network router. This algorithm allows the SeNodeB toemptying its buffer containing the ref ME data before thehandover command, and thus it avoids completely the packetforwarding procedure. The modification in the message chartare reported in Fig. 6, which is related to the time intervalbetween the path switch command and handover command.

The prediction of the handover start-time can be done byusing cross-layer information on the bandwidth available atthe SeNodeB for the ref ME. Such a prediction is accurate toup some hundreds of milliseconds, which is above the typicalRTT, namely the time needed to the SeNodeB to emptyingits buffer. The drawback of this solution is that it is morecomplicated than the one presented in the previous subsection.Moreover, if the prediction is erroneous, it could lead to a totaldisconnection of the ref ME from the LTE network. However,

such an event has a negligible probability of occurrence.

B. Active Queue Management

Active Queue Management (AQM) refers to the controltaking place in core-network routers. There are several mo-tivations for using AQM, such as reduced packet losses, re-duced queueing delay and jitter, improved throughput. Theseimprovements are achieved by using as control input thearrival rate of data packets and the queue size for a particularoutgoing link, and as output a decision on how to mark ordrop packets. The AQM mechanism selects the probabilities formarking and dropping packets. Then, each forwarded packetis marked/dropped with such a probability. Servers will thenreact to these marks and adjust their sending rates.

Several AQM mechanisms are evaluated in [15], with andwithout the use of Explicit Congestion Notification (ECN). InTCP, an option is available for notifying the congestion insidethe network [16]. However, in the tunnel connection, the net-work is configured such that between the serving gateway andthe eNodeB the UDP protocol is used, which does not allowfor ECN notification. It follows that the network configurationis ECN-unaware even if the user transport protocol providessuch a feature.

To overcome the limitation described above, the basic ideais to enable the AQM policy with ECN marking in the core-network router and forward such a marking to the user IPheader located at the end of the tunnel. This solution means toprovide a feature enhancement inside the eNodeB and gatewayto mark/drop the packets depending on the user transportprotocol. The mark is forwarded by the eNodeB to the userIP header only if it carries the ECN-aware sequence (seeFig. 7(a)). Otherwise, when the user IP header handles anECN-unaware TCP the packet will be dropped, as shown inFig. 7(b).

X X X X X X 1 0

TOS field of tunnel IP header

Differentiated Services Code Point ECN

X X X X X X 1 0

TOS field of user IP header

Differentiated Services Code Point ECN

(a) ECN-aware 7→ Mark forwarding.

X X X X X X 1 0

TOS field of tunnel IP header

Differentiated Services Code Point ECN

X X X X X X 0 0

TOS field of user IP header

Differentiated Services Code Point ECN

(b) ECN-unaware 7→ Packet dropping.

Fig. 7. Parameters of the reference ME connection during the handover [16].

VI. SIMULATION STUDY

In this section, we validate the performance improvementsof TCP when LTE handover occurs as achieved by the solutionswe have proposed in Section V. To this aim, we have extendedthe ns-2 simulator.

A. The LTE Simulator

At present ns-2 does not support multiple radio interfaces.It does not have flexible tools for the cross-layer control of

TABLE ICOMMON PARAMETERS OF ALL SIMULATION.

Parameter ValueCross-traffic FTPBitrate of router links 12 Mb/sBitrate of server links 100 Mb/sLatency of wired links 2 mseNodeB queue length 200 packetsReordering EnabledDetach time 30 mseNodeB Scheduling policy WFQIP Packet size 1500 byte

communication systems and, moreover, it is not possible toreuse the existing implementations of the wireless channelbecause they do not work for LTE radio interfaces (they areall ad-hoc solution to the specific problem which they areimplemented for). Therefore, we have implemented an LTEsimulator by using the framework Multi InteRfAce Cross LayerExtension for ns-2 (nsMiracle) [17], which is conceived as aset of dynamic libraries that are loaded to add support formulti-technology and cross-layering. It exploits also a patchthat facilitates the use of dynamic libraries in ns-2 [18], whichallows us to load in the simulator only the necessary modulesand makes the simulation faster.

The TCP version used in our simulations is TCP Reno sinceit is widely used in the Internet. Other versions of TCP, suchas SACK, could lead to better performance [19], but TCPReno plays a significant role in the mobile applications [20].The RTO parameters have been set correctly according to thespecifications RFC2988.

The reference scenario of the simulator is depicted in Fig. 1.In the simulator, the starting of a communication between eachME and server follows the TCP three way handshaking [13]and typically the file transfer is started by the server withoutany request.

In Tab. I we have reported the parameters common for allsimulations. The cross traffic on the core-network router isassumed FTP. The detach time is the interval time of the hardhandover. WFQ is the LTE scheduler used by the eNodeB.Other schedulers give basically same performance results asthose described in the following.

B. Simulation Results

Fig. 8 shows the probabilities of the optimal case fDI, the

window halving case fDW H, and the timeout case fDSS

(seeSection IV). The source cell serves the ref ME. The numberof mobiles in the target cell varies among 2 and 5. A setof 50 simulations was done for every network configuration.Then simulation data are collected according to three differentclasses: The first class is timeout occurrences, in which thenumber of times that the ref ME experiences a RTO event arecollected. The second class is the window halving occurrencesand collects the times that the ref ME experiences a cwndhalving. Finally the last class represents the case in which

Slow Start Window Halving Nothing0

10

20

30

40

50

60

70

80

90

2 ME3 ME4 ME5 ME

Fig. 8. Probabilities of the timeout case fDSS(left), the window halving

case fDW H(middle), and the optimal case fDI

(right) with the standardhandover procedure.

TABLE IISIMULATION PARAMETERS ABOUT FORWARDING AVOIDANCE

MECHANISM.

Parameter ValueTotal simulation time 5 sNodes attached to SeNodeB 1

Nodes attached to TeNodeB 5

Core network router queue size 90 packetscwndMAX 44 ' 64 KBPrediction Step 130 ms

the handover is transparent to the TCP protocol. From Fig. 8,we observe that the RTO expiration is very frequent, whichconfirms that the mechanisms to improve TCP performanceduring the LTE handover are essential. The improvementsoffered by the solutions presented in Section V are discussednext.

1) Forwarding Avoidance: Here we investigate the benefitsof the two forwarding avoidance mechanisms. Tab. II showsthe key parameters used in the simulations. The predictionstep of 130ms gives a reasonable path switch in advance withrespect to the actual handover. The figure was chosen withrespect to the typical RTT (around 100 ms) to empty the pipeof the core-network router.

Figs. 9 and 10 show the congestion window update andthe RTT experienced by the received packets. Figs. 9(a) and10(a) describe the case of standard LTE handover procedure,in which a spurious timeout occurs due to RTT increasing;9(b) and 10(b) are referred to the application of the fast pathswitch procedure, which is only partially able to avoid thespurious timeout event; better performance is achieved by thehandover prediction (Fig. 9(c) and 10(c)), which avoids alwaysthe spurious timeout and causes only a window halving dueto a packet dropping done by the router buffer.

Fig. 11 shows the probabilities of timeout, window halvingand ideal case similar to Fig. 8, but with the difference thatthe handover prediction algorithm presented in in this paper isemployed. We observe a substantial reduction of the slow start

1 1.5 2 2.5 3 3.5 4 4.5 50

5

10

15

20

25

30

35

40

45

50

Time[s]

CWN

D[p

acke

ts]

(a) Standard procedure case.

1 1.5 2 2.5 3 3.5 4 4.5 50

5

10

15

20

25

30

35

40

45

50

Time[s]

CWN

D[p

acke

ts]

(b) Fast path switch case.

1 1.5 2 2.5 3 3.5 4 4.5 50

5

10

15

20

25

30

35

40

45

50

Time[s]

CWN

D[p

acke

ts]

(c) Handover prediction case.

Fig. 9. cwnd for ref ME during the handover. The handover start-time is 2.3 s

1 1.5 2 2.5 3 3.5 40

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

Time[s]

RTT[

s]

(a) Standard procedure case.

1 1.5 2 2.5 3 3.5 40

0.02

0.04

0.06

0.08

0.2

0.12

0.14

0.16

0.18

0.2

Time[s]

RTT[

s]

(b) Fast path switch case.

1 1.5 2 2.5 3 3.5 40

0.02

0.04

0.06

0.08

0.2

0.12

0.14

0.16

0.18

0.2

Time[s]

RTT[

s]

(c) Handover prediction case.

Fig. 10. RTT for ref ME during the handover. The handover start-time is 2.3 s

Slow Start Window Halving Nothing0

10

20

30

40

50

60

70

80

90

2 ME3 ME4 ME5 ME

Fig. 11. Probabilities of the timeout case fDSS(left), the window halving

case fDW H(middle), and the optimal case fDI

(right) when the handoverprediction is enabled.

frequency and an increasing of the window halving frequency,which are beneficial in terms of throughput. This is due tothe reduced timeouts experimented by the ref ME during thehandover thanks to the prediction technique. It is possibleto show that the conclusions derived from the simulationresults reported in Fig. 9, 10, and 11 hold for other networkconfigurations, namely for networks with a different numberof MEs.

2) Active Queue Management: Here we present simulationsto investigate the advantages achieved by using RED queuemanagement instead of the typical drop-tail policy. Tab. IIIshows the parameter settings of the simulator. The REDparameters have been set according to the suggestion givenin [21].

TABLE IIISIMULATION PARAMETERS ABOUT AQM.

Parameter ValueTotal simulation time 10 sNodes attached to SeNodeB 4

Nodes attached to TeNodeB 3

max p 0.1

thresh 20 packetsmaxthresh 65 packetsCore network router queue size 90 packets

TABLE IVACTIVE QUEUE MANAGEMENT SIMULATION RESULTS.

Implemented Link R → SeNodeB Link R → TeNodeBQueue Management Drop Rate [%] Drop Rate [%]

Drop Tail 2.2 3.9RED (Dropping) 2.2 3.6RED (Marking) 2.1 3.3

In Tab. IV we report the probability of dropping packetson the link between the core-network router and the SeN-odeB, and the link between the core-network router and theTeNodeB. Recall that a smaller dropping probability meansa better quality of service for the ref ME. We observe thatexploiting the RED queue management carries to more ben-efits in the link between router and TeNodeB. Indeed, eventhough after the handover such a queue handles four differentTCP connections and then an increased congestion level, thedropping probability is reduced. According to [22] we can say

Default procedure AQM enabled Prediction enabled AQM + Prediction0

100

200

300

400

500

600

700

Del

iver

ed D

ata

[KB]

2 ME3 ME4 ME5 ME

Fig. 12. Amount of data delivered successfully to the ref ME from the startingof the handover until one RTT after the ref ME connects with the TeNodeB.The leftmost bars are referred to the standard handover, the second groupto the handover with AQM, the third group to the handover with prediction,and the rightmost group to the handover with a simultaneous use of handoverprediction and AQM.

that RED allows better performance with a congested networkin which the early dropping works exactly to avoid more dropsdue to the incoming congestion. We conclude that if the crossnetwork ECN is implemented, the marking benefits also canbe exploited, this reduces the drop rate specially for the morecongested link. The conclusions derived from the simulationresults that we have reported in Tab. IV hold also for othernetwork configurations.

Finally, in Fig. 12 we compute the amount of data deliveredfrom the starting of the handover until one RTT after themoment in which the ref ME is connected to the TeNodeBfor each improving solution and a combination of them. Wesee a substantial increase of data delivered, particularly whenthe prediction technique is associated with AQM.

VII. CONCLUSION

In this paper we presented a complete analysis of TCPperformance during the LTE handover. We observed that thethroughput of the mobile equipment doing the handover dropssignificantly. This is due to a forwarding procedure of packetsfrom the source evolved NodeB to the target evolved NodeB.Therefore we proposed and evaluated several improving so-lutions. Specifically, we investigated a forwarding avoidancealgorithms (which includes a prediction technique and a fastpath switch) and active queue management. Extensive numeri-cal simulations, obtained by an extension of the ns-2 simulator,showed that the handover prediction algorithm is the mostpromising technique to improve TCP performance during theLTE handover.

Future work includes an extension of the theoretical analysispresented in this paper to model analytically the TCP flowsand provide an optimization of the buffer size of the router.Furthermore, we plan to study the joint effect of the handoverpredictor and active queue management.

ACKNOWLEDGMENT

The authors are grateful to Andras Racz, Ericsson Research,Budapest, for fruitful suggestions to design the simulator usedin this paper, for providing us a trace file needed for thesimulation implementation, and for the comments for writingthe paper. REFERENCES

[1] K. Bogineni, R. Ludwig, P. Mogensen, V. Nandlall, V. Vucetic, B. Yi,and Z. Zvonar, “LTE Part I: Core network,” Communications Magazine,IEEE, vol. 47, no. 2, pp. 40–43, February 2009.

[2] T. Ye, X. Kai, and N. Ansari, “TCP in wireless environments: problemsand solutions,” IEEE Communications Magazine, vol. 43, pp. 27– 32,March 2005.

[3] R. Lundwig and R. Katz, “Making TCP robust against spurious retrans-mission,” ACM Computer Communication Review, vol. 30, pp. 30 – 36,January 2000.

[4] H. Balakrishnan, S. Seshan, and R. H. Katz, “Improving ReliableTransport and Handoff Performance in Cellular Wireless Networks,”ACM Wireless Networks, vol. 1, no. 4, 1995.

[5] J. Moon and B. J. Lee, “Rate-adaptive snoop: a TCP enhancementscheme over rate-controlled lossy links,” IEEE/ACM Transactions onNetworking, vol. 14, pp. 603–615, June 2006.

[6] M. Ivanovich, P. Bickerdike, and J. Li, “On TCP performance enhancingproxies in a wireless environment,” IEEE Communications Magazine,vol. 46, pp. 76–83, September 2008.

[7] M. Chiang, “Balancing transport and physical layers in wireless mul-tihop networks: jointly optimal congestion control and power control,”IEEE Journal on Selected Areas in Communications, vol. 23, no. 1, pp.104–116, January 2005.

[8] F. Ren, X. Huang, F. Liu, and C. Lin, “Improving TCP Throughput overHSDPA Networks,” IEEE Transactions on Wireless Communications,vol. 7, pp. 1993 – 1998, June 2008.

[9] N. Mller, “Window-based congestion control: Modeling, analysis anddesign,” Ph.D. dissertation, School of Electrical Engineering, KTH,Automatic Control, Osquldas vg 10, 10044 Stockholm, July 2008.

[10] W. Liao, C. Kao, and C.-H. Chien, “Improving TCP performance inmobile networks,” IEEE Transactions on Communications, vol. 53, pp.569– 571, April 2005.

[11] L. Bajzik, P. Horvath, L. Korossy, and C. Vulkan, “Impact of Intra-LTE Handover with Forwarding on the User Connections,” Mobile andWireless Communications Summit, 2007. 16th IST, pp. 1–5, July 2007.

[12] A. Racz, A. Temesvary, and N. Reider, “Handover Performance in3GPP Long Term Evolution (LTE) Systems,” Mobile and WirelessCommunications Summit, 2007. 16th IST, pp. 1–5, July 2007.

[13] J. Postel, “Transmission Control Protocol,” Internet EngineeringTask Force, RFC 0793, Sep. 1981. [Online]. Available: http://www.rfc-editor.org/rfc/rfc793.txt

[14] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA) andEvolved Universal Terrestrial Radio Access (E-UTRAN); Overalldescription; Stage 2,” 3rd Generation Partnership Project (3GPP), TS36.300, Sep. 2008. [Online]. Available: http://www.3gpp.org/ftp/Specs/html-info/36300.htm

[15] L. Le, J. Aikat, K. Jeffay, and F. D. Smith, “The effects of active queuemanagement on web performance,” in SIGCOMM ’03. ACM, 2003.

[16] K. Ramakrishnan, S. Floyd, and D. Black, “The Addition ofExplicit Congestion Notification (ECN) to IP,” Internet EngineeringTask Force, RFC 3168, Sep. 2001. [Online]. Available: http://www.rfc-editor.org/rfc/rfc3168.txt

[17] “ns-MIRACLE: Multi InteRfAce Cross Layer Extension for ns-2.” [On-line]. Available: http://www.dei.unipd.it/ricerca/signet/tools/nsmiracle

[18] “Dynamic modules in ns-2.” [Online]. Available: http://telecom.dei.unipd.it/ns/miracle/ns dynamic libraries/

[19] F. Xin and A. Jamalipour, “TCP throughput and fairness performance inpresence of delay spikes in wireless networks,” Int. J. Commun. Syst.,vol. 18, no. 4, pp. 395–407, 2005.

[20] S. Fu and M. Atiquzzaman, “Modelling TCP Reno with spurioustimeouts in wireless mobile environments,” Oct. 2003, pp. 391–396.

[21] S. Floyd, “RED: Discussions of Setting Parameters,” November 2007.[Online]. Available: http://www.aciri.org/floyd/REDparameters.txt

[22] S. Floyd and V. Jacobson, “Random early detection gateways forcongestion avoidance,” IEEE/ACM Trans. Netw., vol. 1, no. 4, pp. 397–413, 1993.